Issue with grouped operations and attributes

[Enchufa2]

Consider the following custom S3 class with an attribute:

x <- 1:10
class(x) <- "foo"
attr(x, "foo") <- 1:10

in which the length of the attribute must coincide with the vector. I define the following S3 methods:

as.data.frame.foo <- as.data.frame.numeric

mean.foo <- function(x, ...) {
  foo <- mean(attr(x, "foo"))
  structure(NextMethod(), "foo" = foo, class = "foo")
}

where mean summarises the vector as well as the attribute. Then,

data.frame(x=x) %>%
  summarise_all(mean) %>%
  .$x
# [1] 5.5
# attr(,"foo")
# [1] 5.5
# attr(,"class")
# [1] "foo"

Ok, but

data.frame(x=x, y=rep(c("A", "B"), each=5)) %>%
  group_by(y) %>%
  summarise_all(mean) %>%
  .$x
# [1] 3 8
# attr(,"foo")
# [1] 5.5              # <- I'd expect: 3 8
# attr(,"class")
# [1] "foo"

[lionel-]

Aggregation of attributes is a really difficult issue because we don't have any information about how to do it. Some attributes are vectorised, some aren't. I'm surprised that the attributes of the original vector are preserved at all actually.

[Enchufa2]

This could work if the subsetting of grouped data were done by calling R's subsetting operator (at least when there are attributes involved). But even if I define the operator [.foo, the example doesn't work, so I suppose that this is done internally, at C++ level. Any chance of having support for this?

[lionel-]

There's the issue of subsetting and there's the issue of aggregating all the data back to one vector. We need C-level genericity of vectors and we have plans regarding this, but it's not a trivial issue.

[hadley]

This is on our long term to do list.